Agents-A1: a 35B Mixture-of-Experts agentic model that scales the horizon, not the parameters

The problem Agents-A1 targets: capability per parameter, not capability per GPU

Most agentic workloads fail not because the underlying model is too small, but because it cannot hold a task together across dozens of tool calls. The README frames this as agent-horizon scaling, and it splits the idea in two: scaling long-horizon trajectories, and scaling heterogeneous agent abilities. The project's claim is that a ~35B model can approach trillion-parameter behaviour if the training data and the distillation recipe push on horizon length instead of raw parameter count.

The intended audience is fairly specific. It is developers and enterprises who want to integrate a model into their own agent loop, and who care about running it themselves rather than calling an API. The README states the project welcomes developers and enterprises to integrate and try Agents-A1. That is a self-hosting audience, not a managed-service audience. If your constraint is latency on a shared endpoint or a vendor SLA, this repository does not address it.

How the three-stage training recipe produces one deployable student

The mechanism described in the README is a pipeline with three stages. First, full-domain supervised fine-tuning aligns a base model with broad agentic behaviours. Second, domain-level teacher models are trained to capture specialised expertise per domain. Third, a multi-teacher domain-routed on-policy distillation step with what the README calls salient vocabulary alignment transfers knowledge from those teachers into a single student model. The stated goal of that third stage is knowledge-transfer efficiency across domains, and the stated result is six heterogeneous domains unified into one deployable student.

The supporting infrastructure is a long-horizon knowledge-action layer that connects external knowledge, actions, observations and verifier outcomes. According to the README, this produces agentic trajectories with an average length of 45K tokens. That number is the central design claim: the horizon, not the weight count, is what the project scaled. Note the vocabulary alignment detail. It implies the teachers do not share an identical token distribution, so routing and alignment are doing real work rather than being decorative.

The repository layout is consistent with a research release: top-level entries are LICENSE, README.md, assets/, docs/, evaluation/ and scripts/. Evaluation code for selected domains was open-sourced alongside the 35B-A3B model, per the news section.

Getting the Agents-A1 weights and running a first tool-use prompt

The README does not give a pip install line or a Python entry point. It points to distribution channels instead: a Hugging Face collection, a ModelScope model page, a quantised collection, and mlx-community builds for Mac. Treat the model as the artifact and the repository as the training, evaluation and script layer.

Start from the collection the README links, which is where the weights live. The README does not state an exact download command, a repository identifier inside the collection, or a CLI flag, so the first step is to open the collection page and take the model identifier from there rather than guessing one.

For Mac users, the README thanks mlx-community for providing quantised versions at multiple scales and says to try running Agents-A1 on a Mac, which makes the mlx-community collection the practical path on that platform. The README also mentions a series of quantised model variants in the Agents-A1 collection, released on 2026.7.2.

Once weights are local, the repository's scripts/ and evaluation/ directories are where the runnable code lives. The README does not document a single canonical inference command, so the honest first step is to read scripts/ and evaluation/ rather than assume a CLI exists. If you want the smallest thing that runs, the 4B model released on 2026.7.14 is the entry point the README highlights for building a local AI assistant.

What the benchmark table claims, and what it does not settle

The README reports Agents-A1 at 56.4 on Seal-0, 46.4 on HiPhO, 79.0 on FrontierScience-Olympiad, 40.00 on FrontierScience-Research, 80.6 on IFBench and 94.8 on IFEval, calling those overall SOTA. It also reports 75.5 on BrowseComp, 86.0 on XBench-DS-2510, 96.0 on GAIA, 44.3 on SciCode, 47.6 on HLE with tools and 56.8 on MolBench-bind, described as best among comparable models around 35B.

The comparison set is where the table gets interesting. Agents-A1 is placed against Qwen3.5-35B-A3B, Qwen3.6-35B-A3B and Nex-N2-mini in the comparable band, and against Step-3.5-Flash, Kimi-K2.6, DeepSeek-V4-pro (Max) and GPT-5.5 (xhigh) in the larger-scale band. Several entries are marked as wins for the larger models: BrowseComp at 84.4 for GPT-5.5, XBench-DS-2510 at 90.0 for Kimi-K2.6 and DeepSeek-V4-pro, GAIA at 98.06 for DeepSeek-V4-pro. Agents-A1 is not claiming to beat everything. It is claiming to be close, at a fraction of the size.

Read the table with two caveats. First, these are the project's own reported figures, and the README does not describe the harness configuration behind each number in the excerpt available. Second, a benchmark average tells you little about your task. The evaluation/ directory exists precisely so you can run your own. Do that before trusting any single row.

Where Agents-A1 is the wrong tool

The clearest limitation is operational. This is a set of weights plus training and evaluation code, not a serving stack. The README describes integration at the model level and does not document a hosted API, a rate limit, an uptime target or a versioned endpoint. If your product needs a stable remote contract, you are building that layer yourself.

Second, the release history is short and uneven. The 35B-A3B model, evaluation code for selected domains and the technical report arrived on 2026.6.26. Quantised variants followed on 2026.7.2, and the 4B model on 2026.7.14. There are no recent releases retrieved for this repository. A project whose model line moved from 35B to 4B inside three weeks is still settling, and the README gives no deprecation or migration policy.

Third, the 45K-token average trajectory length is a cost, not just a feature. Long horizons mean long contexts, which means memory and latency scale with the task rather than with the prompt. The README presents long-context understanding as a highlight and does not discuss serving cost. For a short classification call, this model is overkill and you will pay for the horizon you are not using.

How Agents-A1 differs from a general-purpose open model like Qwen

Qwen3.5-35B-A3B and Qwen3.6-35B-A3B appear in the README's own comparison table, which makes them the natural reference point. The difference in approach is in what was scaled. A general-purpose instruct model is trained to answer well across a broad distribution of single-turn and short multi-turn requests. Agents-A1's recipe is built around trajectories that chain knowledge, actions, observations and verifier outcomes, then distilled from per-domain teachers into one student.

That shows up as a trade. A domain-routed distillation scheme buys cross-domain competence in one checkpoint, but it also means the model's behaviour is shaped by however the six domains were routed and balanced. The README does not publish the routing weights or the domain mixture. With a general instruct model you get a more predictable, more widely documented behaviour profile and a larger ecosystem of tooling around it. With Agents-A1 you get a model that is explicitly tuned for long tool-use chains, at the cost of less visibility into why it behaves the way it does on an out-of-domain prompt.

Licence, maintenance and what an upgrade actually costs

The repository is licensed Apache-2.0, and the README links to the LICENSE file at the repository root rather than restating terms. Apache-2.0 is a permissive licence with an explicit patent grant, which is generally friendly to commercial integration, but the licence governs the repository. Whether the model weights carry the same terms is something to confirm on the Hugging Face and ModelScope pages, because model cards can carry their own licence. This is not legal advice; read both the LICENSE file and the model card before shipping.

On maintenance, the last push to this repository was on 2026-07-16. That is roughly two months before the date of writing, and the repository is not archived. The README's news section is the only changelog visible, and it stops at the 4B release on 2026.7.14. There is no documented release cadence, no versioning scheme for the weights, and no stated support window. Upgrading means re-downloading a checkpoint and re-validating your prompts against it, because nothing in the README promises behavioural stability across model versions.

Editorial conclusion

Agents-A1 is worth adopting if you need a locally runnable agentic model that handles long tool-use trajectories and you are willing to host a 35B-A3B checkpoint yourself, or to step down to the 4B variant. It is the wrong choice if you need a managed endpoint, a stable long-term API contract, or a model whose numbers you can reproduce without the repository's own evaluation harness. Before committing, verify two things: that the benchmark table in the README matches what you get on your own task, and that the licence file in the repository root is the Apache-2.0 text you expect, since the README links to it rather than restating its terms.

Frequently asked questions

Official sources

1. InternScience/Agents-A1 on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README